The Indian government is dramatically boosting support for domestic artificial-intelligence models in its drive to secure better technologic self-sufficiency, easier access to computing power and AI systems tuned to India’s languages, institutions and economy.
The government, under the newest extension has okayed 20 indigenous AI foundation-model proposals and also granted incentivised computing support to 237 projects. Among the foundation-model calls, proposals have been approved for 12 large multimodal models and eight small language models (Fig. 4b), while approved projects have a total allocation of 93.18 lakh GPU hours.
This is a departure from data merely to spur AI adoption in the world to training LLMs and developing foundational models of AI in India.
India’s sovereign AI strategy
India’s AI strategy is founded on the goal of building custom-made Artificial Intelligence technology that is developed, trained and deployed locally.
In March 2024, the Union Cabinet approved ₹10,371.92 crore for five years towards implementation of the IndiaAI Mission. The task is made to help computing infrastructure, data platforms, native fashions, startups and research and liable AI improvement.
This has a pillar on foundation-model which concentrates on developing large multimodal models, large language models in addition to smaller specialisation-specific models that can perform well given Indian languages and local context.
Indian models would therefore be specific to the linguistic diversity of the country, public-sector requirements etc unlike the general-purpose systems that have been built for broader english-speaking markets.
The bigger picture is not to isolate India in the global AI landscape. Rather, it is to ensure that Indian business houses & researchers and government agencies have domestic alternatives and means of being involved in the development of advanced AI systems.
Twenty foundation-model proposals approved
As part of the IndiaAI Mission, the government has now identified 20 indigenous foundation-model proposals to be funded. The cohort consists of 12 large multimodal models, and eight small language models.
Starting with the large multimodal models that are able to take in and produce mixed forms of information such as text, image, sound and video. They can handle more complicated applications such as document analysis, visual inspection, voice interfaces and automated decision-support systems.
Small language models usually need fewer computing resources and can be easily deployed on local servers, edge devices and mobile hardware. Which makes them very appropriate for applications where cost, privacy and low latency comes into play
The selected projects are expected to address areas such as:
- Indian-language translation.
- Speech recognition and voice assistants.
- Healthcare and medical documentation.
- Education and personalised learning.
- Agriculture and rural services.
- Financial technology and fraud detection.
- Governance and citizen services.
- Defence and cybersecurity.
- Enterprise automation.
- Legal and regulatory analysis.
The government has indicated that intellectual-property rights will remain with the applicants, giving participating companies and institutions an incentive to commercialise their models.
Earlier cohort expands the ecosystem
The latest approvals build on an earlier IndiaAI foundation-model programme.
Earlier, twelve organisations and consortia were identified to create both large- and small-scale language models based on Indian datasets. Examples include Sarvam AI, Soket AI, Gnani AI, Gan AI, BharatGen consortium (led by IIT Bombay), GenLoop, Zenteiq, Intellihealth and Shodh AIs from M S Swaminathan Research Foundation and techstack Partners collaborate with Fractal Analytics & Tech Mahindra Maker’s Lab.
The organisations selected are an eclectic mix of start-ups, trading tech businesses and academia. Intended to foster competition, this combination is also designed to link research institutions with commercial deployment capabilities.
According to the government, the frameworks from which these models are created must be beneficial for the open-source ecosystem and released publicly so they can be used by both government organizations as well as startups and researchers.
The greater the public access, the less chance there is — over time — that India’s AI market became completely a captive of a few large international platforms. It would also enable developers to develop specific applications without needing to renegotiate with international model vendors.
Subsidised compute becomes central
Foundation model training requires huge compute resources. Developers require large datasets, high-performance computing clusters with thousands of GPUs connected via high-speed networking and storage.
To this end the IndiaAI Mission has developed a compute-access programme in partnership with the public and private sector. Impaneled 15 compute service providers and approved 237 projects for subsidised AI computing All these projects together contributed 93.18 lakh GPU hours.
The capability of the mission has also been increased on GPUs. The IndiaAI compute ecosystem, offering one of the largest such clouds in the world, was targeting or onboarding over 38,000 GPUs to democratize access to high-performance computing for startups, researchers and public institutions by the government.
That is useful since the computing costs can hinder smaller organisations from effectively training or tuning advanced models. The absence of subsides means Indian startups will either have to cut their models down to more compact and less effective systems or rely on pricey foreign cloud platforms.
Full subsidy for core model development
The government has also announced a complete subsidy on computing infrastructure expenses incurred in the process of building foundational AI models. Core model development receives the full subsidy, while vertical AI applications and inference workloads receive a smaller subsidy.
The objective of the policy is to overcome one of the biggest constraints that Indian AI companies have: The cost of training a model before it starts making any money.
For a startup, the cost structure can be divided into several stages:
- Data collection and cleaning.
- Model design and experimentation.
- Pre-training.
- Fine-tuning.
- Safety testing.
- Deployment and inference.
- Monitoring and maintenance.
The subsidy relates mostly to the pre-training and development phase which can prove to be very expensive. The reality is, companies still need to pay for staff & datasets & product development, compliance / customer support and commercial operations.
This structure incentivizes companies to focus public support on the capital- and technology-heavy stage, but forces them to develop sustainable business models in all remaining stages of the product lifecycle.
Why Indian-language models matter
India has hundreds of languages and dialects, but many of the leading AI systems perform much better in English than in Indian languages.
Models in Indian languages can make technology more accessible to those not as comfortable with English as the dominant online language and who might prefer Hind, Tamil, Telugu, Bengali, Marathi, Punjabi, Kannada, Malayalam, Gujarati and Odia.
Possible applications include:
- Voice-based government services.
- Regional-language education platforms.
- Local-language banking assistance.
- Automated translation for courts and public offices.
- Agricultural advisory systems.
- Healthcare communication in rural areas.
- Customer support for small businesses.
- Voice interfaces for low-literacy users.
The problem is not how to translate, Language. AI systems need to recognize local cultural references, social context, legal jargon, accents and dialects, mixed-language speech.
Like Indian users tend to write his sentence mix with English and any regional language. Models that are trained on this sort of real-world usage might be more useful than systems that were design around formal written language solely by definition.
Foundation models versus application tools
A foundation model is a large AI system trained on broad datasets and capable of supporting many different tasks. Application companies can build products on top of these models for specific sectors.
For example, an Indian-language foundation model could be used to create:
- A healthcare assistant.
- A tax-support chatbot.
- A legal research platform.
- A government translation tool.
- A customer-service agent.
- A financial document analyser.
This eventually similar to the relationship of an operating system and applications, which is in a layered structure. The foundation model supplies the general intelligence and the startups provide specific interfaces and workflows around them.
India policy is the only option with two layers of support. The government is backing core models but providing subsidised computing for the startups and researchers who build applications.
BharatGen and academic participation
One of the most ambitious academia-led efforts in India is the BharatGen consortium led by IIT Bombay and supported under the IndiaAI Mission.
BharatGen aims to build the generative AI models suited for Indian languages and contexts, with participation from academia and research institutions. It is an attempt by the government to establish a research-led alternative to commercial AI development.
Academic involvement is important for several reasons:
- Universities can conduct long-term research.
- Researchers can publish methods and benchmarks.
- Public institutions can develop open datasets.
- Students can be trained in advanced AI engineering.
- Models can be designed for public-interest applications.
However, academic projects must also address deployment challenges. A model that performs well in research testing may still require major improvements in speed, safety, reliability and operating cost before it can serve millions of users.
Support for startups and private companies
Indian startups will likely play a big role in turning publicly supported research into commercial products.
Selected participants such as Sarvam AI, Gnani AI, Gan AI and others are working on language, speech, multimodal and enterprise applications.
Private-sector participation brings several advantages:
- Faster product development.
- Better understanding of customer needs.
- Commercial distribution networks.
- Ability to recruit specialised talent.
- Experience in cloud deployment.
- Access to international markets.
Startups also face greater pressure to demonstrate revenue and user adoption. Public funding can reduce development risk, but it cannot replace customer demand.
The most successful companies are likely to combine public support with private investment, strategic partnerships and long-term enterprise contracts.
Strengthening India’s compute infrastructure
The foundation-model exercise is heavily intertwined with India’s larger strategy to develop indigenous capabilities in AI.
The government is helping the development of public and private GPU clusters, cloud platforms and data centres. This compute portal of the mission is designed to match AI developers with computing resources available for subsidised access.
Local compute capacity can provide several benefits:
- Lower latency for Indian users.
- Better control over sensitive data.
- Reduced dependence on overseas providers.
- Greater access for startups and universities.
- Support for models trained on domestic datasets.
- Improved resilience during international supply disruptions.
The cost of building and operating the computing infrastructure is, however, high. GPU clusters need reliable power, advanced cooling, high speed networks and specialized technical staff.
So India will have to scale up not just its GPU inventory but also its power, data-centre and semiconductor capabilities.
AI safety receives greater attention
The government’s revised approach also has a stronger focus on responsible and safety-related AI.
We have supported Projects related to AI safety, evaluation and responsible development. These initiatives aim to mitigate risks associated with harm, bias, misinformation, privacy and cybersecurity issues; hallucinations; and harmful Automated Decision Making.
When it comes to public sector products like social media, finance, healthcare and education systems especially, safety is a critical component of AI.
One that creates false information in a friendly chat might be inconvenient. In a medical, legal or financial system, the same mistake could have disasterous effects.
Important safety measures include:
- Testing for language and demographic bias.
- Evaluating factual accuracy.
- Protecting personal and sensitive data.
- Monitoring harmful or manipulated content.
- Creating human-oversight procedures.
- Publishing performance benchmarks.
- Building systems for reporting and correcting errors.
India’s foundation-model programme could become more credible if participating organisations publish clear evaluation results and document how their models handle safety risks.
The challenge of data
Training Indian-language models requires large quantities of high-quality data.
Some datasets may potentially be split between public government archives, printed books and websites (internet), audio recordings of these events, educational materials or private databases. A lot of it will still need cleaning, translation, annotation or legal vetting.
India has Unveiled its new Government AI-Kosh platform to help access datasets & public digital assets intact with Public Resources required for the development of AI. The IndiaAI Mission views data platforms and access to compute as integral components of the national AI ecosystem.
As much as the quantity of data, it equally becomes crucial to maintain the quality of data. Even if you have powerful computer resources at your disposal a model trained on incorrect, redundant or biased data will generate unreliable output.
One must include copyright, privacy and consent as well which developers also have to deal with. Even though information is publicly available, it does not mean that it is free from the constraints of law or ethics.
India’s competitive position
So government intervention is necessary as also this global race for the most advanced AI models.
Both the United States and China remain by far the leaders in model development, computing capacity and in private investment in AI. European or Asian economies are also launching national strategies for reducing their dependency on foreign platforms.
India may not be looking for the biggest model in the world. In the Indian language rich and priority sector orientated ecosystem, it maybe more practical to develop efficient & focused systems at optimum inputs which work best there.
India could gain an advantage through:
- Multilingual AI.
- Affordable inference.
- Small models for mobile and edge devices.
- Public-sector applications.
- Digital public infrastructure.
- Healthcare and agriculture use cases.
- Large engineering and developer communities.
- Open-source and collaborative development.
A smaller model that is cheaper to operate and better suited to a specific Indian workflow could be more commercially valuable than a much larger general-purpose model.
Risks and limitations
The expanded mission also faces significant challenges.
Funding concentration
If most support goes to a small number of organisations, the ecosystem may not develop enough competition. Smaller startups and academic groups could struggle to access compute, data and talent.
Compute dependency
India may still rely on imported GPUs and international semiconductor supply chains. Domestic model development does not automatically create complete technological independence.
Commercial viability
Foundation models are expensive to train and maintain. Companies must find sustainable revenue through APIs, enterprise licences, government contracts or specialised applications.
Model performance
Indian-language models may perform well in selected benchmarks but struggle with reasoning, factual accuracy, long-context understanding or complex multilingual conversations.
Data governance
Unclear rules around copyright, privacy and public data could slow model development or create legal disputes.
Talent scarcity
India has a large software workforce, but fewer specialists with experience in training frontier-scale models, designing AI infrastructure and doing advanced safety research.
Evaluation standards
Without independent and transparent benchmarks, it may be difficult for users to compare Indian models with international alternatives.
Implications for businesses
The mission could create new opportunities for Indian companies across the technology sector.
Businesses may gain access to lower-cost AI models for:
- Customer support.
- Content generation.
- Document processing.
- Search and knowledge management.
- Business analytics.
- Translation.
- Voice automation.
- Coding assistance.
- Cybersecurity.
- Personalised education.
Affordable Indian-language AI can lower the cost of automation for small and medium enterprises.
Other benefits to Digital publishers and web business could include the creation of content tools in local-language, automating metadata generation, relying on speech-to-text systems, or translation and regional search optimisation. However, businesses must not rely on AI-generated content without verification and they should ensure that human review is still in place for all factual/legal/financial material.
Impact on developers
Indian developers could gain access to public APIs, open-source models, datasets and subsidised GPU capacity.
This may encourage the development of applications specifically designed for Indian users rather than adapting products created for foreign markets.
Potential developer opportunities include:
- Regional-language SaaS tools.
- Voice-first mobile applications.
- AI assistants for government schemes.
- Indian legal and compliance software.
- Financial-literacy platforms.
- Agricultural advisory applications.
- Education technology.
- Healthcare documentation.
- Small-business automation.
- AI tools for local retailers and service providers.
The availability of domestic models may also improve control over data and reduce dependence on third-party foreign APIs.
From policy support to practical results
The success of the IndiaAI Mission will ultimately depend on what happens after models are funded and trained.
The government will need to ensure that selected projects meet measurable milestones, including:
- Development of usable model prototypes.
- Transparent performance testing.
- Support for multiple Indian languages.
- Demonstrated safety and reliability.
- Access for startups and researchers.
- Deployment in real-world public and commercial settings.
- Sustainable operating costs.
- Regular updates and maintenance.
The programme should also encourage interoperability. If models, datasets and APIs are developed using open standards, developers will be able to combine technologies from different organisations.
India’s AI mission enters a new phase
With the expansion of the IndiaAI Mission, the government is moving beyond high-level policy statements to direct implementation of essential elements for AI development.
It has approved 20 indigenous foundation-model proposals, supported for 237 subsidised-compute projects and allocated93.18lakh GPU hours, thereby providing Indian startups, researchers and institutions with an enhanced access to advanced technology enabling them to compete in the AI space.
The success of the programme will not be simply a count of models. That will depend on whether those models are effective, affordable, secure, multilingual and useful in real Indian contexts.
India need not mimic everything that Silicon Valley has done (or other AI powers). Its best opportunity may be to focus on building efficient systems for Indian languages, public infrastructure, local industries and large-scale social applications.
Thus, if the state marries public finance and commitment to evaluation with private-sector talent, ethics in data governance and scalable computing infrastructure, the IndiaAI Mission may be an anchor point of a truly unique Indian AI ecosystem.